ShebMichel commited on
Commit
66fb36d
·
verified ·
1 Parent(s): 14fad6b

Upload agent_trace.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. agent_trace.json +74 -0
agent_trace.json ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "project": "Bedtime Story Machine",
3
+ "track": "Thousand Token Wood",
4
+ "builder": "ShebMichel",
5
+ "tools_used": [
6
+ "Kiro CLI (AI agent)",
7
+ "Python",
8
+ "Gradio",
9
+ "Modal",
10
+ "ffmpeg"
11
+ ],
12
+ "models": {
13
+ "story_generation": "nvidia/Nemotron-Mini-4B-Instruct (4B params)",
14
+ "illustration": "black-forest-labs/FLUX.1-schnell (12B params)",
15
+ "total_params": "~16B"
16
+ },
17
+ "infrastructure": {
18
+ "inference": "Modal (Nemotron on T4 GPU)",
19
+ "image_gen": "HF Inference API",
20
+ "hosting": "Hugging Face Spaces",
21
+ "frontend": "Gradio with custom CSS"
22
+ },
23
+ "build_steps": [
24
+ {
25
+ "step": 1,
26
+ "action": "Designed app architecture - chose Nemotron-4B for text, FLUX.1-schnell for images"
27
+ },
28
+ {
29
+ "step": 2,
30
+ "action": "Built Gradio app with storybook UI (custom CSS, alternating scene layout)"
31
+ },
32
+ {
33
+ "step": 3,
34
+ "action": "Implemented story generation with structured JSON output + regex fallback"
35
+ },
36
+ {
37
+ "step": 4,
38
+ "action": "Integrated FLUX.1-schnell for watercolor-style illustrations"
39
+ },
40
+ {
41
+ "step": 5,
42
+ "action": "Added French language support via language dropdown"
43
+ },
44
+ {
45
+ "step": 6,
46
+ "action": "Set up Modal deployment for Nemotron-4B with vLLM"
47
+ },
48
+ {
49
+ "step": 7,
50
+ "action": "Created demo video with ffmpeg subtitle burning"
51
+ },
52
+ {
53
+ "step": 8,
54
+ "action": "Deployed to HF Spaces under build-small-hackathon org"
55
+ },
56
+ {
57
+ "step": 9,
58
+ "action": "Published blog post and agent trace"
59
+ }
60
+ ],
61
+ "challenges": [
62
+ "Nemotron-4B not available on any free inference provider - solved with Modal deployment",
63
+ "HF token needed fine-grained Inference API permissions",
64
+ "JSON parsing reliability from small models - added regex extraction fallback",
65
+ "Consistent illustration style - solved with prompt engineering suffix"
66
+ ],
67
+ "badges_targeted": [
68
+ "\ud83d\udfe9 NVIDIA Nemotron Quest (Nemotron-4B model)",
69
+ "\ud83d\udc1c Tiny Titan (\u22644B story model)",
70
+ "\ud83d\udfe2 Modal Awards (Modal deployment)",
71
+ "\ud83d\udcd3 Field Notes (blog post)",
72
+ "\ud83d\udce1 Sharing is Caring (this trace)"
73
+ ]
74
+ }